Learning from limited echocardiographic data requires representations that capture both cardiac anatomy and motion. This study combines Higher Order Dynamic Mode Decomposition (HODMD) [1] with a Vision Transformer pretrained as a Masked Autoencoder (MAE) [2, 3] for patient-level classification into Normal, Ischemic and Hypertrophic groups. HODMD extracts spatial modes and their temporal frequencies from the echocardiographic sequences [4] . Three illustrative loops yield cardiac-periodic patterns and accurate heart-rate estimates. For classification, we compare three routes: MAE frame features alone, frame features augmented with a 16-dimensional modal spectral vector, and frame features combined with five modal magnitude maps processed by a shared encoder. The study uses 679 clips from 30 patients, with ten patients per class. All routes start from the same MAE checkpoint, pretrained on unlabelled frames from the entire cohort. Supervised evaluation uses a fixed, stratified five-fold patient partition, repeated over ten training seeds. Adding modal magnitude maps increases mean patient-level accuracy from 0.693 to 0.747 and macro-F1 from 0.693 to 0.743. The spectral-vector route provides no improvement in either mean metric. These results suggest that spatial modal maps complement frame-based MAE features and show promise for improving classification on limited data.